CLEAR: Cross-Layer Extreme Activation Representation for Backdoor Sample Detection
Year: 2026
Type: preprint
Abstract: The application of deep learning models in security-critical scenarios makes trustworthy training data essential for model reliability. However, existing backdoor sample detection methods often fail to capture subtle anomalous responses induced by stealthy triggers, particularly under low poisoning rates, complex trigger patterns, and the absence of clean validation data. To address these limitations, this paper proposes CLEAR, a Cross-Layer Extreme Activation Representation framework for backdoor sample detection. First, batch normalization (BN) statistics align feature distributions across hidden layers, while channel-wise extreme responses are extracted as backdoor anomaly representations to amplify subtle trigger-induced activation differences. Second, to avoid relying on prior knowledge for layer-weight selection, particle swarm optimization models hidden-layer attention weights as particle positions and uses the stability of HDBSCAN clustering after Uniform Manifold Approximation and Projection (UMAP) as the fitness function, adaptively determining the optimal weight combination that maximizes the anomalous consistency of poisoned samples. Finally, density-based clustering is applied to optimally weighted features, and a predefined cluster-size threshold identifies the most stable and compact anomalous micro-cluster, enabling prior-free screening of poisoned samples without clean labels. Experiments on CIFAR-10 and Tiny-ImageNet show that CLEAR achieves average true positive rates (TPRs) of 97.54% and 99.54%, respectively, with average false positive rates (FPRs) of 0.16% and 0.03%, significantly outperforming AC, SCALE-UP, AIBD, and CCD. Moreover, models retrained on datasets purified by CLEAR maintain classification accuracy close to that of benign models while achieving substantially lower attack success rates (ASRs), demonstrating superior detection accuracy, robustness, and overall defense performance.
Source: Research Square
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225780006
Cites:
Cited by:
Related to: